Skip to content

lcmgregoire/UPDATE_KAR_PIPELINE_LEASYSCAN_EZTr

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Update of Kar et al 2020 pipeline

This project concerns an update of the pipeline described in the article :Kar, S., Tanaka, R., Korbu, L.B., Kholová, J., Iwata, H., Durbha, S.S., Adinarayana, J., Vadez, V., 2020. Automated discretization of 'transpiration restriction to increasing VPD' features from outdoors high-throughput phenotyping data. Plant Methods 16, 140. https://doi.org/10.1186/s13007-020-00680-8

Link GITHUB to the original pipeline : https://github.com/KSoumya/EZTr

The script developed by Kar et al. 2020 concerns the analysis of data from the HTP Leasyscan platform. This platform consists of :

  • .... load cells able of measuring weights at 15min intervals.
  • PlantEye scanner for monitoring architecture/ plant growth traits (especially leaf area) . The scanner is capable of taking 2 measurements per day.
  • Weather data monitoring, at 15-minute intervals.

This script here is based on the analysis of "Exp60" carried out at ICRISAT, Hyderabad, India, in September-October 2023. The measurement period is from September 27 to October 17, with 6 irrigation days removed. Leaf area measured by PlantEye measuring 3D-LA leaf area twice a day is supplemented by leaf area data observed with a planimeter at the end of the trial. The plant material is composed of 320 sorghum genotypes of RefSet list defined by the Genbank at ICRISAT. During the analysis of this experiment, we observed some data outliers that were not cleaned, and aberrants ETref values. We therefore decided to take a closer look at the filtering stages and identified certain steps where modifications would be necessary. We present below the modifications we made to the initial script, followed by some simple suggested improvements. We also noticed that the pipeline was adapted to the specific case of the ICRISAT platform. A number of enhancements are envisaged as possible improvements (not yet present in the script) to adapt it to other HTP platform (different LC size, different location...)

Modifications made to the initial version

1) Outlier detection in the curateRawLC package

Initial version: detection of weight outliers by boxplot and negative values.

However, if a LC contains a majority of outliers, the data of this LC will not be removed. For exemple, data of LC "141:03:01" with a mean at 200 kg on the boxplot, instead of a +/- 70kg expected weight, will not be removed.

image

Current version: you defined an expected weight parameter (avg_wgt, in grams) at the beginning of the pipeline. If the weight is outside a range of +/- 30% of this expected weight, it is considered an outlier.

2) ETref calculation in calculateETref package

Initial version: ETref values too large for the expected order of magnitude. We identified a conversion problem of the solar radiation value in the initial version. In addition, ETref calculation is adapted to a 15 min interval only.

Current version: the conversion problem is fixed and ETref calculation is suitable for all measurement intervals (defined by the seq_by parameter).

3) PlantEye :

Initial version: There is no visualization of leaf area data measured by the Plant Eye scanner to check whether the 3D-LA data used to calculate Tr and TRrate are consistent. However, Plant Eye can provide unreliable or missing data when plants have reached a certain phenological stage (barcode reading problems, overlapping, etc.). This is why, when we observe the leaf area profile over time, we see leaf area growth at the start of the experiment, followed by an inflection point, then followed by missing or too low data in many cases.

Current version: we propose here to visualize roughly the 3D-LA data of all LC over time, and you can remove by yourself the measurements days when you consider 3D-LA values as outlier (The aim being to keep only the first linear part of the 3D-LA profile.)

image

Once you have confirmed your choice, you can always correct it by looking at the plots of LAI per day/LAI per Timestamp or LA per Timestamp of each LC later on the process. (example here where Oct 10-11, 2023 were kept at first sight, then deleted when viewing the LAI/LA graphs).

1LAI_over_time

1LAI_over_time

4) Management of missing LAI/LA data:

Initial version: for missing LAI data (often at the end of the experiment), the pipeline calculates the median LAI on the days measured and assigns this value to all missing data. The problem is that this value is therefore lower than the last days measured, and that this value is a constant for the missing days. It therefore does not represent leaf growth.

image

Current version: with Plant Eye data that the you have chosen to keep, previously converted to observed LA (+ planimeter data if available), the pipeline calculates LA regression as a function of time, assuming linear regression. The slope value is recovered, and the missing values are assigned the previous day's value + slope value.

5) Tr calculation in the calculateTr package:

Tr calculation follows this formula detailled in Kar et al.

image

Initial version: Tr= 1-(1-exp(-0.463*LAI)))*ETr

Current version: Tr= (1-exp(-0.463*LAI)))*ETr

6) TRrate calculation in the calculateTr package:

Initial version: This part is in comment and therefore not used.

Current version :

image

as described in Kar et al.2020

Improvements proposed

The main priority for us was to work on data visualization. Indeed, in the initial version of the pipeline, we have just the final outputs. So, we have no visual verification of the filtering process, no clues as to whether the imputed data are logical, whether the transpiration profiles are consistent, etc..

1) You can plot the results of an LC to see the outlier filtering process, and check that the profiles obtained are correct.

Ex: Weight profile before and after filtering

image image

Ex : Evapotranspiration profile before and after filtering

image image

2) Visualise data one LC:

  • Leaf area over time (LA, in m²)

1LA_over_time

  • Evapotranpsiration (ETr non normalized, mm.15min-1) and Reference Evapotranspiration (mm.15min-1) to check if ETr is lower than ETref

1_ETr_before_norm

  • Transpiration (Tr, mm.15min-1)

1_Transpiration

  • Transpiration and ETr (mm.15min-1) to check if Transpiration is lower than Evapotranspiration

ETr_1_ETr_and_Transp

  • Transpiration rate (TRrate, mm.15min-1.m-²)

1_TRrate

3) Save/load objects created in the process (folder /saved_objects/)

4) Outlier detection :

Count of negative and max values before and after filtering processes

5) In the generateETrRatio package of the initial version

ETr filtering according to ETref value is based on an hourly system to define the diurnal (6:30am-6:30pm) and nocturnal (6:30pm-6:30am) periods. This works in the case of India, but cannot be exported for other conditions. We have therefore chosen to base this filter on solar radiation.

  • If SR<0  night time  ETr= 0
  • If SR>0  daytime  -0.01<ETr<ETref

6) Filtering ETr/TRrate according to profile

After filtering, we observed that some high peaks remained on the ETr profile normalized by LA or TRrate. The weather data were similar from one day to the next, that did not allow us to determine the cause of this difference. This does not concern a particular day. It affects many LCs, but not all. We decided to retrieve the peak max for each day of measurement and make a boxplot with the max values. If a value is considered an outlier on the boxplot, then the entire day of measurement is removed.

Improvements to come:

The pipeline is adapted for Leasyscan platform in India for a few crops. We would like to adapted it to other HTP platform, to other crops. To reach that goal, it will be necessary to work on the following points :

  1. ET conversion from grams to mm is based on the ICRISAT LC area = ¼ m² (in the convETr package)

  2. LAI and LA calculations (and therefore affect the value of Tr and TRrate) are based on the LAI = 0.26, adapted to the ICRISAT LC ( in the calculateTr package).

  3. Conversion calculations from the 3D value estimated by PlantEye and the value observed on the planimeter.

The equations to convert 3DLA in observed LA come from the article Vadez et al 2015. These equations are based on three crops: pearl millet, cowpea and peanut. In the initial version of the pipeline, the equation is correct in the case of pearl millet, incorrect in the other cases (cowpea and peanut). It would be wise to i) correct the equation for both crops, ii) make a loop so that you can decide which crop corresponds best to yourcase, iii) In addition, it would be interesting to have more crops proposed in order to refine the relationship.

  1. ...

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages